Reproduce “Shallow AutoEncoding Recommender with Cold Start Handling via Side Features” on OpenResearch

About this paper

Edward DongBo Cui, Lu Zhang, William Ping-hsun Lee

Researchers from DotDash Meredith and NBC Universal Peacock developed FEASE, an extension of the EASE model, to address both user and item cold start problems in recommender systems by integrating side information. The method effectively recommends to new users and items while preserving performance for existing ones, leveraging EASE's computational efficiency.

Key insights

• The computational efficiency and closed-form solution of the EASE model can be successfully extended to handle cold start problems by integrating side information without sacrificing its core benefits.
• Explicitly incorporating both user and item side information is crucial for robust cold start performance, as traditional EASE methods assign near-zero recommendation scores to cold items.
• A tunable trade-off exists between maximizing overall recommendation accuracy for warm items and effectively improving recommendations for cold items, which FEASE allows practitioners to manage.

Summary by alphaXiv.